How yiwenai/autosat works
Project Type and Purpose
• AutoSAT is an open-source tool (Multi-Agent Version) for automatically optimizing heuristics in SAT solvers using Large Language Models (LLMs). It orchestrates multiple agents (advisor, coder, evaluator) to generate, execute, and evaluate heuristic functions against SAT instances.
High-Level File Structure Analysis
• Top-level:
• Dockerfile, setup.py, requirements.txt – build/deployment support
• main_MultiAgent.py & evaluate.py – CLI entry points for training and testing
• autosat/ – core Python package
• data/ – scripts and templates to generate CNF instances
• examples/ – sample configs, prompts, and solver code
• autosat/ subpackages:
• agent/ – agent base classes and implementations (assistant_agent, executor_agent, common_agent)
• llm_api/ – abstraction over LLM API calls
• execution/ – execution_worker that compiles/runs SAT solvers with generated heuristics
• evaluation/ – evaluate.py for computing PAR-2, #solved, etc.
• trainer/ – orchestrates multi-agent training loops
• utils.py – shared utilities
Key Insights from README
• Two main workflows:
a. Training via main_MultiAgent.py with multi-agent loops (iteration_num, batch_size, agent_type, etc.)
b. Evaluation via evaluate.py (runs final heuristics on a test set)
• Data generation scripts live in data/ and create CNF instances for training/evaluation
• Examples folder holds prompts, configs, and baseline solver code (EasySAT variants)